When I heard that Bradley Efron and Trevor Hastie had released a new book, I was duly excited. Ever heard of the bootstrap? That's just a little something Efron proposed a couple of decades ago, that has since become about as ubiquitous as the normal distribution (just kidding, but seriously). Have you ever read any of the two Data Science bibles, Introduction to Statistical Learning (ISL) or Elements of Statistical Learning (ESL)? Well, that's Hastie, along with another of my heroes Rob Tibshirani (father of the LASSO) and Jerome Friedman (inventor of gradient boosting).
Are you getting excited?
So this book sums up about 60 years of developments in statistical methods, and it's written by the very people driving those developments. Overall, the book is very well written, with concise language, useful examples, consistent and clear terminology and mathematical notations, and a great overarching perspective on a big and quickly evolving field.
One of the unique things with this book is the historical perspective. The authors don't only describe the methods that were enabled by powerful computing in the recent decades, they also discuss the historical links between them and their philosophical underpinnings. Statistical methods never appear in a vacuum, they build on top of previous work and are guided by the contemporary schools of thought.
Something I really appreciated while reading Computer Age Statistical Inference was the conceptual division between frequentist, Bayesian and Fisherian theory. In the beginning of the book, these contrasting philosophies are described and throughout the book, they are used to understand methods like regularization, boosting and bagging, and even neural networks. This has made a lot of things more clear to me. Also, after finishing the book, I started to understand that these theories aren't necessarily mutually exclusive - they just put emphasis on different aspects of statistical inference. The best way forward may be to reconcile and integrate thinking from all three theories, like the authors argue in the last chapter about Empirical Bayes methodology.
Even though I read the entire book and put effort into following the text and understanding the models, I don't feel like I'm done with this book. I will definitely go back and read the chapters again and use it as a reference. And I'm sure I will understand a little more every time I do it.
Check it out at the official website.